Model Drift
Degradation of a deployed model's predictive accuracy over time as input feature distributions or outcome relationships shift from the training environment.
FAQs
- How quickly can model drift make a pricing model materially inaccurate?
- The timeline varies widely by line of business and the pace of environmental change. A model may remain stable for years in a static environment or degrade significantly within months following a major external disruption such as a pandemic, a litigation environment change, or a macroeconomic shift. Continuous monitoring is more reliable than fixed revalidation calendars for catching rapid drift.
- Is retraining always the right response to detected drift?
- Not necessarily. Retraining on recent data corrects for gradual covariate shift but may cause the model to overfit to a transient anomaly if the drift reflects a temporary disruption rather than a permanent structural change. The response should be determined by the type and source of drift, informed by subject matter expert judgment about whether environmental changes are likely to persist.
- Should we monitor vendor models for drift the same way we monitor internally built models?
- Yes. You are responsible for model outcomes regardless of who built the model. Require vendors to provide monitoring metrics and drift reporting as part of the service agreement, and build independent monitoring on your end using the predictions and inputs you observe in production.
Related Terms
MLOps Insurance
Practices adapting machine learning operations to insurance: model versioning, deployment pipelines, monitoring, retraining, and regulatory documentation.
Model Governance
Policies, controls, and oversight processes managing the full lifecycle of predictive and AI models from development through retirement.
Feature Engineering
Selecting, transforming, and constructing input variables from raw data to improve predictive accuracy of machine learning models in insurance.
Claims Severity Model
A model predicting the ultimate cost of an individual claim, used to set reserves, prioritize handling, and flag high-exposure files.
